23 September 2024 in IEEE Transactions on Intelligent Transportation Systems
In recent years, the surge of artificial intelligence has propelled autonomous driving technology to the forefront, capturing growing interest and enthusiasm. As a sub-module within autonomous driving, research in traffic sign detection and recognition has significantly advanced with a growing focus on utilizing deep learning methods. Nevertheless, in complex real-world road scenarios, traffic signs often exhibit a long-tailed distribution, with the majority of instances concentrated in a few frequent categories and a scarcity in the remaining ones. Considering conventional Traffic Sign Detection and Recognition methods are crafted using manually curated datasets, the class imbalance may detrimentally impact the efficacy of the detection model. In this paper, we first propose a gradient-guided loss reweighting model that dynamically reweights the loss for positive and negative samples based on the cumulative gradients across each category. Additionally, a classification bias-based refinement module is proposed to fine-tune these weights during training, based on the false positive and false negative rate. This serves to suppress a drop in precision for tail categories resulting from the gradient-guided loss reweighting module, thus further balancing the entire training process for improved results. Extensive experiments are performed on the TT100K and GTSDB datasets, yielding significant advancements surpassing state-of-the-art methods.
7 August 2024 in IEEE Sensors Journal
Road surface condition is a crucial factor affecting vehicle safety. Accurate and rapid identification of road conditions can enhance the adaptability of active safety control systems to adapt to complex working conditions, while predicting the road adhesion coefficient in advance can help active safety control systems adjust control strategies promptly to ensure driving safety and stability. This paper introduces an algorithm that combines the squeeze-and-excitation network (SENet) and convolutional neural network (CNN) to estimate the road adhesion coefficient based on the reflectance intensity features of LiDAR point clouds. The method mainly comprises the following seven steps: (i) data acquisition; (ii) point cloud segmentation; (iii) selection of the point cloud region of interest (ROI); (iv) filtering processing; (v) grayscale image generation; (vi) classification based on the SE-CNN model; and (vii) combination of the graph method to estimate the road adhesion coefficient. Data were collected on urban roads to construct the dataset. Experimental results indicate that different lighting conditions and speeds have minimal impact on the distribution of reflectance intensity in regional point clouds. The classification accuracy of this research method for dry asphalt, wet asphalt, dry concrete, and wet concrete reached 100%, 99.98%, 99.82%, and 99.23%, respectively. Compared to other machine learning methods, the proposed method exhibits higher accuracy and stability in predicting road surface adhesion coefficients.
14 February 2025 in Journal of Lightwave Technology
The performance of received signal strength (RSS)-based visible light positioning (VLP) is heavily reliant on the pre-existing channel model which describes the RSS attenuation with distance. In indoor areas, dust particles accumulating on the surfaces of LEDs may cause the pre-existing channel model deviating from actual situations and consequently decrease the localization performance. In this paper, we first experimentally verify the impact of dust on the RSS VLP system performance and find that the positioning error is significantly increased even in slight dust deposition scenarios. Then we modify the classical Lambertian-based channel model to reflect the influence of dust, and propose a two-points approach which performs the channel model calibration only based on two RSS measurements. Additionally, the variations of the channel model parameters with the accumulation of dust are observed and mathematically modeled. A single-point approach is further designed which requires only one RSS measurement for each channel model calibration. The experimental results demonstrate that the average positioning error remains below 6 cm under different dust accumulation scenarios. The negative impact of dust can be effectively mitigated by our proposed RSS VLP scheme with low system maintenance cost. To the best of our knowledge, this is the first piece of work considering dust accumulation for VLP.
1 January 2022 in Journal of Central South University
5 March 2025 in Transportmetrica B: Transport Dynamics
20 December 2024 in Neural Networks